NVIDIA: Llama 3.3 Nemotron Super 49B V1.5

Description

Llama-3.3-Nemotron-Super-49B-v1.5 is a 49B-parameter, English-centric reasoning/chat model derived from Meta’s Llama-3.3-70B-Instruct with a 128K context. It’s post-trained for agentic workflows (RAG, tool calling) via SFT across math, code, science, and multi-turn chat, followed by multiple RL stages; Reward-aware Preference Optimization (RPO) for alignment, RL with Verifiable Rewards (RLVR) for step-wise reasoning, and iterative DPO to refine tool-use behavior. A distillation-driven Neural Architecture Search (“Puzzle”) replaces some attention blocks and varies FFN widths to shrink memory footprint and improve throughput, enabling single-GPU (H100/H200) deployment while preserving instruction following and CoT quality. In internal evaluations (NeMo-Skills, up to 16 runs, temp = 0.6, top_p = 0.95), the model reports strong reasoning/coding results, e.g., MATH500 pass@1 = 97.4, AIME-2024 = 87.5, AIME-2025 = 82.71, GPQA = 71.97, LiveCodeBench (24.10–25.02) = 73.58, and MMLU-Pro (CoT) = 79.53. The model targets practical inference efficiency (high tokens/s, reduced VRAM) with Transformers/vLLM support and explicit “reasoning on/off” modes (chat-first defaults, greedy recommended when disabled). Suitable for building agents, assistants, and long-context retrieval systems where balanced accuracy-to-cost and reliable tool use matter.

Raw fields snapshot

{
  "id": "nvidia/llama-3.3-nemotron-super-49b-v1.5",
  "name": "NVIDIA: Llama 3.3 Nemotron Super 49B V1.5",
  "description": "Llama-3.3-Nemotron-Super-49B-v1.5 is a 49B-parameter, English-centric reasoning/chat model derived from Meta’s Llama-3.3-70B-Instruct with a 128K context. It’s post-trained for agentic workflows (RAG, tool calling) via SFT across math, code, science, and multi-turn chat, followed by multiple RL stages; Reward-aware Preference Optimization (RPO) for alignment, RL with Verifiable Rewards (RLVR) for step-wise reasoning, and iterative DPO to refine tool-use behavior. A distillation-driven Neural Architecture Search (“Puzzle”) replaces some attention blocks and varies FFN widths to shrink memory footprint and improve throughput, enabling single-GPU (H100/H200) deployment while preserving instruction following and CoT quality.\n\nIn internal evaluations (NeMo-Skills, up to 16 runs, temp = 0.6, top_p = 0.95), the model reports strong reasoning/coding results, e.g., MATH500 pass@1 = 97.4, AIME-2024 = 87.5, AIME-2025 = 82.71, GPQA = 71.97, LiveCodeBench (24.10–25.02) = 73.58, and MMLU-Pro (CoT) = 79.53. The model targets practical inference efficiency (high tokens/s, reduced VRAM) with Transformers/vLLM support and explicit “reasoning on/off” modes (chat-first defaults, greedy recommended when disabled). Suitable for building agents, assistants, and long-context retrieval systems where balanced accuracy-to-cost and reliable tool use matter.\n",
  "created": 1760101395,
  "canonical_slug": "nvidia/llama-3.3-nemotron-super-49b-v1.5",
  "hugging_face_id": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5",
  "source_type": "openrouter_only",
  "context_length": 131072,
  "max_completion_tokens": null,
  "is_moderated": false,
  "architecture": {
    "modality": "text->text",
    "input_modalities": [
      "text"
    ],
    "output_modalities": [
      "text"
    ],
    "tokenizer": "Llama3",
    "instruct_type": null
  },
  "input_modalities": [
    "text"
  ],
  "output_modalities": [
    "text"
  ],
  "modality": "text->text",
  "tokenizer": "Llama3",
  "instruct_type": null,
  "supported_parameters": [
    "frequency_penalty",
    "include_reasoning",
    "max_tokens",
    "min_p",
    "presence_penalty",
    "reasoning",
    "repetition_penalty",
    "response_format",
    "seed",
    "stop",
    "temperature",
    "tool_choice",
    "tools",
    "top_k",
    "top_p"
  ],
  "default_parameters": null,
  "per_request_limits": null,
  "top_provider": {
    "context_length": 131072,
    "max_completion_tokens": null,
    "is_moderated": false
  },
  "pricing": {
    "prompt": "0.0000001",
    "completion": "0.0000004"
  },
  "PPM": {
    "prompt": 0.1,
    "completion": 0.4
  },
  "openrouter_raw": {
    "id": "nvidia/llama-3.3-nemotron-super-49b-v1.5",
    "canonical_slug": "nvidia/llama-3.3-nemotron-super-49b-v1.5",
    "hugging_face_id": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5",
    "name": "NVIDIA: Llama 3.3 Nemotron Super 49B V1.5",
    "created": 1760101395,
    "description": "Llama-3.3-Nemotron-Super-49B-v1.5 is a 49B-parameter, English-centric reasoning/chat model derived from Meta’s Llama-3.3-70B-Instruct with a 128K context. It’s post-trained for agentic workflows (RAG, tool calling) via SFT across math, code, science, and multi-turn chat, followed by multiple RL stages; Reward-aware Preference Optimization (RPO) for alignment, RL with Verifiable Rewards (RLVR) for step-wise reasoning, and iterative DPO to refine tool-use behavior. A distillation-driven Neural Architecture Search (“Puzzle”) replaces some attention blocks and varies FFN widths to shrink memory footprint and improve throughput, enabling single-GPU (H100/H200) deployment while preserving instruction following and CoT quality.\n\nIn internal evaluations (NeMo-Skills, up to 16 runs, temp = 0.6, top_p = 0.95), the model reports strong reasoning/coding results, e.g., MATH500 pass@1 = 97.4, AIME-2024 = 87.5, AIME-2025 = 82.71, GPQA = 71.97, LiveCodeBench (24.10–25.02) = 73.58, and MMLU-Pro (CoT) = 79.53. The model targets practical inference efficiency (high tokens/s, reduced VRAM) with Transformers/vLLM support and explicit “reasoning on/off” modes (chat-first defaults, greedy recommended when disabled). Suitable for building agents, assistants, and long-context retrieval systems where balanced accuracy-to-cost and reliable tool use matter.\n",
    "context_length": 131072,
    "architecture": {
      "modality": "text->text",
      "input_modalities": [
        "text"
      ],
      "output_modalities": [
        "text"
      ],
      "tokenizer": "Llama3",
      "instruct_type": null
    },
    "pricing": {
      "prompt": "0.0000001",
      "completion": "0.0000004"
    },
    "top_provider": {
      "context_length": 131072,
      "max_completion_tokens": null,
      "is_moderated": false
    },
    "per_request_limits": null,
    "supported_parameters": [
      "frequency_penalty",
      "include_reasoning",
      "max_tokens",
      "min_p",
      "presence_penalty",
      "reasoning",
      "repetition_penalty",
      "response_format",
      "seed",
      "stop",
      "temperature",
      "tool_choice",
      "tools",
      "top_k",
      "top_p"
    ],
    "default_parameters": null,
    "expiration_date": null
  }
}